Related Experiment Video
Updated: Jul 9, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
DSGH-Net: Medical Image Segmentation via Dual-Statistical Dynamic Context and Graph-Convolutional Heterogeneous
Jiying Ma1, Di Long1, Pengfei Li2
1College of Computer Science and Technology; Liaoning Provincial Key Laboratory of Intelligent Technology for Chemical Process Industry, Shenyang University of Chemical Technology, No. 11 Street, Shenyang, 110142, Liaoning, China.
Journal of Imaging Informatics in Medicine
|July 7, 2026
Summary
DSGH-Net enhances medical image segmentation by improving context-adaptive modeling and hierarchical feature decoding. This novel network achieves superior performance on datasets like Kvasir-SEG, demonstrating robust generalization capabilities.
Area of Science:
- Medical image analysis
- Computer vision
- Machine learning
Background:
- Accurate medical image segmentation is vital for clinical diagnosis and treatment planning.
- Current CNN- and Transformer-based methods struggle with variations in lesion size, morphology, and ambiguous boundaries.
- Existing models face limitations in context-adaptive modeling and balancing deep semantic and shallow edge feature recovery.
Purpose of the Study:
- To propose DSGH-Net, a novel network for medical image segmentation.
- To address challenges in context-adaptive modeling and hierarchical feature decoding.
- To improve the accuracy and robustness of automatic medical image segmentation.
Main Methods:
- Introduced the Dual-Statistical Context Modulation Block (DCM-Block) for dynamic multi-scale feature fusion.
- Developed the Heterogeneous Stage Decoder (HSD) with Topological Dynamic Context Refinement Block (TDCR-Block) and Semantic Calibration Graph Convolution (SC-GCN) for deep stages, and a lightweight CNN for shallow stages.
- Incorporated the Deep Weighted Fusion Attention (DWFA) module in skip connections for enhanced feature propagation.
Main Results:
- DSGH-Net achieved competitive performance on Kvasir-SEG and ISIC2018 datasets, outperforming existing methods.
- Achieved a Dice score of 92.48% and IoU of 87.80% on the Kvasir-SEG dataset.
- Demonstrated strong cross-dataset generalization, achieving a Dice score of 87.63% when transferred from Kvasir-SEG to CVC-ClinicDB.
Conclusions:
- DSGH-Net effectively addresses limitations in context-adaptive modeling and hierarchical feature decoding for medical image segmentation.
- The proposed network shows significant improvements in accuracy and robustness across different medical imaging datasets.
- DSGH-Net offers a promising solution for challenging medical image segmentation tasks, with potential for clinical applications.